Skip to content

Vibe coding glossary

By Mari Luukkainen · 15 min read · Last updated

On this page

A vibe coding glossary is a list of the words you meet when you build software with AI tools. This one defines 59 terms from A to Z in plain language, including prompt, agent, Git, commit, deploy, API key, RLS and worktree. Each term has its own link, so you can send someone straight to one definition.

Definitions describe the words as they are used in vibe coding, not their full technical meaning. Other guides on this site link to these terms the first time they use them, such as prompt, commit and deploy.

Ask AI about this pageChatGPTClaudePerplexityGoogle AI ModeGrok

Terms from A to Z

AGENTS.md
A project instruction file read by several coding agents, including Codex, Claude Code, GitHub Copilot, OpenCode and Cursor. It holds the same kind of content as CLAUDE.md: commands, folder layout and rules. Write the rules once in AGENTS.md, and give tools that want their own file a one-line import such as @AGENTS.md.
AI coding agent
A tool that takes a task, reads your project, edits files, runs commands such as tests and reports back, often asking permission before it acts. Claude Code, Codex, OpenCode and Antigravity CLI are examples. An agent differs from a chat assistant because it works directly on your files instead of only giving you code to paste.
API
A defined way for one program to ask another for something, such as weather data or a payment. Your app sends a request to an API and gets a response. Most APIs require a key to identify you. Keep that key on the server, not in browser code.
API key
A secret string that identifies your app to a service and may allow charges to your account. Keep it as private as a password. Never put it in frontend code or commit it to Git. If a key is exposed, replace it, because deleting the commit does not make it secret again.
Authentication
Checking who someone is, usually with a login. It answers the question of identity. Building it yourself is risky, so most apps use a managed login service. Test it by trying to reach a protected page in a private window while signed out.
Authorization
Checking what a signed-in person is allowed to do or see. It is separate from login. A common AI-written flaw is a page that shows the right data for your ID, but also for anyone else's if they change the ID in the address.
Backend
The part of an app that runs on a server and not in the browser. It handles saving data, logging in, sending email and talking to other services. Secret keys and permission checks belong here. Beginners often skip it for a first project, and add it when the app needs accounts.
Branch
A separate line of work in Git. You try a risky change on a branch, and if it works you merge it into the main line, and if it fails you delete the branch. Branches let you experiment with an AI tool without risking the version that currently works.
Build
The step that turns your source code into the files that are published. Running npm run build locally before you deploy catches many problems, because the host runs the same step. Fix every error the build reports without switching its checks off.
Checkpoint
A saved state of your files that a coding agent records before each prompt, so you can restore it. In Claude Code, pressing Esc twice or typing /rewind opens the menu to restore code, the conversation or both. Checkpoints cover the agent's own file edits, not changes made by shell commands, so keep using Git.
CI
Continuous integration, an automatic run of your tests, lint and build every time you push code. A red result tells you about a problem before the change reaches your published app. Pull requests from coding agents usually show the CI result, so check it before you merge. CI does not replace reading the diff.
CLAUDE.md
A text file in a project root that Claude Code reads at the start of each session. Put the commands to run and test the project and a short list of rules in it. Other tools use similar files. See how to use Claude Code.
Commit
A saved snapshot of your project in Git, with a short message describing the change. Commit after each working round, before you ask the AI for something big. If the next prompt breaks the project, you can discard the uncommitted changes and return to the last commit.
Context window
The amount of text an AI model can use at once, including your messages, the files it has read and its own replies. In a long session older instructions can fall out or lose weight. Start a fresh session for each task to keep rules effective.
--dangerously-skip-permissions
A Claude Code startup flag that runs the agent without asking permission for file edits and commands. Anything it decides to run, including deleting files, happens without a prompt. Use it only in a disposable environment such as a throwaway container with no secrets, never in a folder with data you cannot rebuild.
Database
A program that stores your app's data in an organised way, such as users, orders or posts, and lets the app search and change it. A first project often needs none. When you add one, the main risk is access: check that one user cannot read another user's rows.
Dependency
A package of code written by someone else that your project uses. Dependencies save time and add risk, since you rely on code you did not write. Ask your AI tool not to add dependencies without asking, and check that a named package exists and is still maintained.
Deploy
To publish your app so other people can open it at a web address. With a modern host you push your code to GitHub and the host builds and publishes it automatically. Always repeat your test list on the published version, because it can behave differently from your computer.
Dev server
The copy of your app that runs on your computer while you build, usually started with npm run dev. It reloads when files change. A message that the address is already in use means an old copy is still running. Close it or stop it, then start again.
Diff
A display of what changed between two versions of your files, with removed lines and added lines marked. Reading the diff is how you review what an AI did. If a one-line request changed twelve files, read the diff closely before you commit.
Docker
A way to package an app with everything it needs to run into a container, so it behaves the same on your computer and on a server. Hosts that build from a Dockerfile use it. You rarely need to write one yourself: ask your AI tool to create it and explain each line.
.env file
A plain text file in your project that holds environment variables, one per line. It must be listed in .gitignore so it is never committed. Tell your AI tool in its rules file never to read or print it. Create a new .env at each place the app runs.
Environment variable
A named setting, such as a key or a web address, kept outside your code and read when the app runs. Locally it lives in a .env file. At the host you set it in the project settings. Environment variables keep secrets out of Git and let one codebase run in different places.
Framework
A ready-made structure for building a kind of app, which supplies routing, page layout and many common features so you write less. Next.js is a common one for websites. AI tools tend to choose a framework for you when you do not name one, so state a preference if you have one.
Frontend
The part of an app that runs in the visitor's browser: the layout, buttons and what they see. It is where most beginner projects start. Anything placed in frontend code, including secret keys, can be read by a visitor, so sensitive work belongs on the backend.
Git
A tool that records snapshots of your project so you can compare versions and return to an earlier one. It is what makes mistakes recoverable in vibe coding. Run git init once in a project folder, then commit after every round of work that functions properly.
GitHub
A website that hosts Git repositories and adds issues, pull requests and collaboration tools. Most hosting services can publish an app directly from a GitHub repository, so pushing your code there also starts your deployment. You need a free account to follow most beginner guides.
Hallucination
When an AI states something false with confidence, such as a function that does not exist, a package that was never published or a setting that is wrong. Guard against it by running the code, running tests and checking that named packages are real before installing them.
Hook
A command that a coding agent runs automatically at a set moment, such as after every file edit or before a shell command. A hook that runs the linter and type check after each edit catches errors while the agent is still working. Unlike a line in CLAUDE.md, a hook runs every time.
Hosting
The service that keeps your app running and reachable on the internet. Static sites need very little. Apps with accounts and saved data need a server and a database. Look up a host's plan limits on its own page before launch, because they change and a busy app can exceed them.
Hydration
In React and Next.js, the step where the browser attaches interactivity to HTML the server already produced. A hydration error means the server's HTML and the browser's first render differ, often because the page reads browser-only data such as localStorage too early. Paste the full error into your tool and say what you changed.
Lint
An automatic check that flags suspicious code and style problems, run with a command such as npm run lint. A clean lint result does not mean the app works, but it catches mistakes early. Tell the AI to fix lint errors, not to turn the rule off.
LLM
Large language model, the kind of AI system behind Claude, GPT and Gemini. It predicts text from the text it has seen, which is how it writes code, and it can state wrong things with confidence. Coding tools wrap an LLM with file access and commands. The model behind a tool affects results and cost.
localhost
The address of your own computer. When the dev server runs, you open it at http://localhost:3000 or a similar number. Only you can see it. To let others see your app you deploy it to a host. A localhost address in a published link is always a mistake.
MCP
Model Context Protocol, a standard way to connect AI tools to outside programs and data, such as a database or a documentation search. An agent with MCP connections can use those tools during a task. Each connection widens what the agent can reach, so add only what you need.
Next.js
A framework for building websites and web apps with React. It handles pages, routing, server code and search engine basics in one project. Many AI tools choose it by default for web projects. Its error messages are a frequent subject of beginner questions.
Node.js
A program that runs JavaScript outside the browser. Most modern web projects need it to run on your computer, and it includes npm. If an AI tool says it cannot run your project, ask it to check whether Node.js is installed and to explain how to install it.
npm
The package manager for JavaScript projects. Commands such as npm install download dependencies and npm run dev starts your app. The file package.json lists what your project uses and which commands exist. Most beginner errors involving npm come from incompatible dependency versions or from running it in the wrong folder.
Plan mode
A way of working where the AI tool lists what it intends to do and waits for your approval before editing anything. It is useful for any change bigger than a small fix, because it exposes a wrong direction in seconds. You can get the same effect by asking for a plan in your prompt.
Prompt
The text you give an AI tool to describe a task. A good prompt names the outcome, the limits on what may change and a check the tool can run. Vague prompts get guesses. The prompt library has more than thirty you can copy and adapt to your own project.
Prompt injection
An attack where text the AI reads, such as a web page or a document, contains instructions that the AI then follows. It matters when your app or agent reads content from strangers. Limit what the AI is allowed to do with that content, and never give it secrets.
Pull request
A proposal to merge a branch into the main line on GitHub, with the changes shown for review. People read the diff, comment and approve. Coding agents can open pull requests for you. Reading the diff before merging is the review step that catches mistakes.
Rate limit
A cap on how many requests a person or program may make in a period. On your app it stops bots from flooding a form or guessing passwords. On a service you call, it limits how fast you can use it. AI-written forms often lack one, so ask for it.
Refactor
Changing the structure of code without changing what it does, to make it easier to understand and change. Do it in small steps with tests running before and after. A project edited by prompts for weeks usually needs a refactor, because duplicated logic builds up.
Repository
A project folder tracked by Git, with its full history of changes. People say repo for short. A repository can sit on your computer, on GitHub or both. Keeping a project in a repository means you can go back to any earlier saved state when a change breaks something.
RLS
Row Level Security, a database feature that decides which rows each user may read or change. In Supabase it must be switched on for each table and needs a policy, such as rows visible only to their owner. With it off, anyone holding the public key can read the table. CVE-2025-48757 describes an insufficient policy in Lovable.
Rules file
A text file that tells an AI tool how to work in your project, for example CLAUDE.md, AGENTS.md or a Cursor rules file. Keep it short: commands, folder layout and a few rules that came from real mistakes. It saves you from repeating the same instructions in every prompt.
Skill
A packaged set of instructions that a coding agent loads only when a task needs it, instead of carrying them in every session. Examples are a checklist for publishing a release or a house style for tests. Skills keep the rules file short, since detailed procedures load when they are relevant.
Slash command
A command typed at a coding agent's prompt starting with a forward slash, such as /init, /clear, /compact or /rewind in Claude Code. Type a slash to see the list. Slash commands manage the session itself: they clear the conversation, switch the model, show permissions or restore an earlier checkpoint.
Static site
A site made of finished files served as they are, with no server code running per visitor. It is fast, cheap to host and has few security risks. Landing pages, blogs and small tools like the recipe scaler in how to vibe code can be static.
Subagent
A separate AI instance, started by your main coding agent, with its own context window and a narrow job, such as reviewing a diff or searching the codebase. It reports a result back instead of filling the main conversation. A reviewer subagent should not be the one that built the change.
Supabase
A hosted database service built on Postgres that also provides login, file storage and an automatic API. Many tutorials use it as the default back end, and Lovable Cloud builds on Supabase's open source foundation. Its protection depends on Row Level Security rules, so check that every table has them before you launch.
Terminal
A text window where you type commands to your computer. On macOS it is the Terminal app, on Windows PowerShell. Terminal agents such as Claude Code run inside it. You do not need to learn many commands: a handful for moving between folders, running the app and using Git cover most of a beginner's needs.
Test
A small program that checks that part of your app behaves as intended. Tests give an AI tool something real to check its work against. A failing test written before the fix defines what fixed means. A passing test proves only what it checks.
Token
A small chunk of text that a language model reads or writes, roughly a word or part of one. Usage limits and context size are counted in tokens. You do not need to count them, but long sessions and large pasted files use more of your allowance.
TypeScript
JavaScript with types, which flags mistakes such as passing text where a number is expected before the app runs. Many AI tools choose it by default. Type errors in the build output look intimidating, but pasting the full message into your tool usually produces a fix.
Vercel
A hosting company for web apps, made by the creators of Next.js. You connect a GitHub repository, and every push is built and published automatically. It also runs v0, an AI app builder. Add environment variables in its project settings before the first deploy, and check the plan limits before launch.
Vibe coding
Building software by describing what you want to an AI in plain language, running the result and telling the AI what to fix. The term was coined by Andrej Karpathy in February 2025. Your work shifts from typing code to describing, testing and deciding. See what is vibe coding for the full explanation.
Worktree
A second checked-out copy of a Git repository in a separate folder, on its own branch, sharing the same history. It lets two coding agents work on two tasks at once without overwriting each other's files. Create one with git worktree add and remove it with git worktree remove after you merge.

Frequently asked questions

Which terms should a beginner learn first?

Prompt, Git, commit, deploy, API key and environment variable. They cover starting, saving, publishing and keeping secrets safe.

Which terms matter before I let an agent run without supervision?

Commit, checkpoint, permission prompts, worktree and RLS. Together they decide whether you can undo what the agent did and whether it can reach data it should not.

What do I do when a term is not in the glossary?

Ask your coding tool: Explain this term in two sentences, and show where it appears in this project. An answer tied to your own files is easier to remember than a general definition.

Next guideHow to vibe code →